On 19 August 2026, Xero announced that its JAX agent would save accountants around half of their bank reconciliation time; in September, Odoo 20 shipped AI agents that can create and update records. Yet a survey of 486 accounting professionals, published on 10 September 2026, shows that only one in five firms using AI can measure a clear return. AI in accounting works well for repetitive volume; the decisions, the year-end close and the signature remain yours.
Key points
- Vendors mainly promise automatic bank reconciliation and agents that do the data entry for you. The gains they quote are their own figures, not independent measurements.
- In practice, AI is used first for drafting emails, summarising and research: bank reconciliation comes near the bottom of the list (24% of users).
- Professionals do not trust it without review: only 19% would use its output with limited checking.
- Under Swiss law, every entry must be backed by a voucher and remain traceable, and the annual report is signed by people, not by software.
- The right question is not "can AI keep my books?" but "which tasks should it take on, under which rules and with what control?".
What is happening
Vendors are speeding up. At Xero, the JAX agent automatically matches "high-confidence" transactions to bank feeds, explains why it made each match and leaves exceptions to a human. In its 19 August 2026 release, the company claims around 50% savings on accountants' monthly reconciliation time, more than 100 million transactions reconciled since launch, and a new month-end agent. When the beta launched in November 2025, the stated goal was to reconcile more than 80% of bank statement lines automatically, with automation switched on account by account.
Odoo 20, released in September 2026, goes the same way: AI agents that create and update records, an assistant that answers questions about your finances "based solely on the accounting data" in your database, and a command to rerun automatic reconciliation. Two details matter as much as the new features: every AI feature now consumes paid credits, and every entry that touches a bank account must now originate from a bank transaction.
Practitioners are more cautious. The software vendor Financial Cents surveyed 486 accounting professionals in North America between 15 July and 7 August 2026. 95% use or are exploring AI, but only 11% have embedded it in daily work across much of the firm. Among users, 96% rely on a general assistant such as ChatGPT; the leading uses are drafting emails (75%), summarising documents or meetings (71%) and research (69%). Bank reconciliation concerns only 24% of them. One in five reports a measurable return on investment; 52% call it "promising but hard to quantify".
The most telling part lies elsewhere: 90% believe human judgment matters more, not less, as AI spreads, and only 19% trust its output enough to settle for limited review. Their top concern, raised in 38% of open answers, is accuracy, including errors stated with complete confidence. These figures come from vendors who sell to accountants: they show a trend, not settled truth.
AI in accounting: what it changes for you
What AI does well. Reading a PDF invoice and extracting the amount, date and number; matching a statement line to a voucher; applying the same rule a hundred times in the same way; spotting a duplicate or a missing document. That is exactly what vendors target, and it is where the evening hours are won back.
What it does not do. It cannot tell whether a purchase is business or private, whether an invoice that looks like a duplicate is actually a replacement, or how to treat a currency difference nobody has defined. It can suggest; someone must have written the rule, or make the call.
Responsibility cannot be handed to software. In Switzerland, the Code of Obligations requires every entry to be supported by an accounting voucher and entries to remain traceable (art. 957a), and books and vouchers to be kept for ten years (art. 958f). In a company limited by shares, setting the accounting principles is a non-transferable duty of the board of directors (art. 716a). The annual report, prepared within six months of the financial year end, is signed by the chair of the governing body and by the person responsible for financial reporting within the company (art. 958). No agent signs on your behalf.
The year-end close is still a matter of judgment. Accruals, provisions, valuations, tax choices: these are decisions, not matches. AI can prepare the ground: everything reconciled, each voucher attached to its entry, anomalies listed. Your fiduciary or chartered accountant does not disappear; they receive cleaner data, earlier, and can spend their time advising rather than keying in.
What I have learned
Since September 2026, I have kept my own company's books in Odoo with an AI agent. I described it in detail in AI bookkeeping in practice: my company's books in Odoo; here I keep only what answers a business owner's question.
- The agent handles the volume. It reads statements and vouchers, finds invoices in emails, matches them and prepares the entries. From January to October 2026, every bank line of my company was reconciled.
- The decisions are written down, and they are mine. Accounting date equals payment date, a single exchange-rate entry per period, private purchases through a transit account, no expense without a voucher unless I explicitly agree.
- Nothing is written without a safety net. A plan, a dry run, a backup, then execution. When something is off (exchange rates frozen after the migration, a replacement invoice that looked like a duplicate, an amount debited that differed from the invoice), the agent flags it; I make the call.
My conclusion matches that of the professionals surveyed: AI saves time on volume, but safeguards matter more than the model. I did not need a turnkey "accounting agent": an open ERP, written rules and a supervised agent are enough for a small business.
What to do in practice
1. List your repetitive tasks
Importing statements, entering supplier invoices, reconciling, hunting for vouchers, sending reminders. Write down how long they take each month: without a baseline you cannot measure any gain, which is precisely what half of the firms surveyed are missing.
2. Write your rules before automating
Accounting date, exchange rates, purchases paid privately, expenses without a voucher: one page is enough. In the survey, 35% of firms with a written AI policy report a measurable return, against 14% of those that have none and plan none.
3. Keep AI inside your systems
Pasting statements or invoices into a consumer assistant takes your data out of your control and leaves no trace in your books (see shadow AI in the workplace). Prefer the AI built into your software, or an agent connected to your ERP with defined permissions and a log of what it does (see connecting AI to your ERP).
4. Require validation and a trail
Every automatic match must be visible and reversible, every entry must carry its voucher, and any new entry must first go through a dry run or an approval. That is what keeps your books traceable, with or without AI.
5. Involve your fiduciary from the start
Ask them in what form they want the data, what they will check at year end and what they are happy for AI to prepare. And count the real cost: built-in AI is paid by usage, in credits in the case of Odoo 20.
Frequently asked questions
Can AI keep all of a small business's books?
No. It can prepare most of the day-to-day volume, but decisions, the close and the signature remain human, and the law requires vouchers and a trail you must be able to show.
Do I still need a fiduciary or an accountant?
Yes, for the close, tax, advice and, where applicable, the audit. What changes is what they receive: reconciled, documented books instead of a box of receipts.
Should I switch on my software's AI or connect my own agent?
Built-in AI is easy to switch on, paid by usage and limited to what the vendor has planned. An agent connected to the ERP is more flexible, but needs rules, safeguards and someone to keep them up to date.
Going further
Deciding which accounting tasks to hand to AI in your company, and which to keep, is a matter of diagnosis: see the Strategy and guidance page. When the tool has to be built (statement import, connector, agent with safeguards), that is custom applications and automation. To find out where to start, let's talk during a 20-minute call.
Sources
- The State of AI in Accounting & Bookkeeping 2026, Financial Cents (software vendor), published 10 September 2026; online survey from 15 July to 7 August 2026.
- Xero Announces New AI Innovations for Small Businesses and Accountants at Xerocon US, Xero, press release, 19 August 2026.
- Get a first look at automatic bank reconciliation, Xero blog, 20 November 2025.
- Odoo 20 Release Notes, Odoo, September 2026.
- Swiss Code of Obligations (SR 220), art. 716a, 957a, 958 and 958f, Fedlex, Swiss Confederation (unofficial English translation), accessed 10 November 2026.
Photo: Cht Gsml / Unsplash (unsplash.com/photos/FVwy7PBiSUo).
Tags: AI in accounting, accountant, fiduciary, bank reconciliation, year-end close, Odoo, automation